Identification and prediction of the degree of multidimensional returning to poverty risk for the household in China through the novel hybrid model: Based on the survey data of China Family Panel Studies (CFPS).

In China, absolute poverty has been effectively eliminated, but this does not signify the complete eradication of poverty. Instead, poverty persists in the forms of relative and secondary poverty. More concerningly, regions or households lifted out of poverty continue to face numerous risks of returning to poverty. In this context, measuring poverty solely based on monetary metrics is no longer adequate. Furthermore, the primary focus of grassroots governance has shifted from merely assessing poverty to accurately predicting the multifaceted risks associated with falling back into poverty. A multidimensional poverty indicator system is constructed to measure and predict the risk of multidimensional returning to poverty in China. Then, the Alkire-Forster counting method is applied to measure the risk of multidimensional poverty return and demonstrate the contribution of various indicators to multidimensional poverty return using tracking survey data from the China Family Panel Studies (CFPS). The results show that the multidimensional poverty return in China is mainly caused by the factors in two or three dimensions, where the social development capability dimension has the highest contribution to the multidimensional poverty return index with 43.12 %, followed by the health and education dimensions. Moreover, according to the finding that the identification of multidimensional poverty varies with the values of the poverty cut-off, a novel and practicable method is proposed to classify the risk into three levels, noted as the high, medium and low levels. Consequently, a hybrid model is constructed to predict the risk of multidimensional poverty return by integrating the Archimedes optimization algorithm (AOA), variational mode decomposition (VMD) and Bi-directional long short-term memory (BiLSTM) neural networks. Finally, the performance of the constructed model is validated with an accuracy up to 99.81 %. The constructed model's efforts outperform the traditional BiLSTM and several prevalent machine learning algorithms through extensive comparative experiments. These results illustrate that the proposed model can accurately and stably predict the potential risk of poverty return for multidimensional poverty groups and regions. In conclusion, drawing from the analysis of factors contributing to the risk of relapsing into poverty in this study, policy suggestions have been formulated focusing on education, social capacity enhancement, and healthcare system advancement. In summary, this paper provides new insights into the factors contributing to multidimensional poverty recurrence and the methods for assessing its risk levels, while also introducing a more precise approach to predicting these levels.